hierarchical clustering

Forming an investment portfolio is one of the main concerns of managers and investors who strive in order to create the best investment portfolio to get the best return from the market. So far, many methods have been presented to construct a portfolio, of which the most famous is the Markowitz approach. Our research aims to offer a classical portfolio selection using cluster analysis. We trained four models using k-means clustering with daily log returns as features and agglomerative clustering methods with complete, single and average linkages based on correlation-based distances. Four equally weighted portfolios of 30 stocks each were formed by selecting the stock with the highest Sharpe ratio from each cluster. Based on the silhouette scores and Sharpe ratio, we selected agglomerative clustering with average linkage trained on last year’s data as our final model. The performance of our selected portfolios over the test period was better than random selection in terms of Sharpe ratio but worse than the overall index. The results in terms of volatility showed better performance; our selected portfolio had an annualized volatility lower than the random selection and the average volatility of all clusters and relatively close to that of the equally weighted portfolio consisting of all 334 stocks in the data.